← Back to feed

The timing could not have been worse for Nvidia. On June 1, Jensen Huang spent the Computex keynote in Taipei launching Cosmos 3, calling it “the world’s first fully open foundation model for physical AI” and noting it ranked first among open models on major global leaderboards. One day later, a two-year-old startup in Hangzhou took that top spot.

Spirit AI’s Spirit v1.6 scored 1,924 on RoboArena, displacing Nvidia’s Cosmos3-Nano-Policy at 1,881. A second Nvidia-linked entry, DreamZero, placed third at 1,763. Spirit v1.6 also outperformed Physical Intelligence’s pi0.5.

What RoboArena Measures

RoboArena is not a simulated leaderboard. It was developed through collaboration between UC Berkeley, Stanford, and Nvidia, with research selected for an oral presentation at CoRL 2025. It evaluates generalist robot policies on real-world execution: squeezing containers without bursting them, closing drawers, picking unfamiliar objects, planning multi-step tool usage. The benchmark has been called the “Chatbot Arena for robotics” because it breaks single-laboratory evaluation and uses real task completion rather than preference votes.

Nvidia’s Cosmos 3 failed to complete an opening-laptop task in head-to-head video comparisons. Spirit v1.6 completed the same task smoothly and then moved on to a multi-step manipulation sequence involving recognition, grasping, and placement.

Capital Behind the Model

Spirit AI announced the RoboArena result alongside a 1.5 billion yuan ($222 million) Series A+ financing on June 3 — its fourth round in three months. Cumulative fundraising since February has reached approximately 5 billion yuan ($739 million), a pace described as the most aggressive in the embodied intelligence sector.

The April round saw Lei Jun and Jack Ma jointly invest for the first time in the same embodied AI company. The latest round drew top-tier US dollar funds, industrial capital, and state-backed investors. Spirit AI’s valuation now exceeds 20 billion yuan ($2.9 billion).

Deployment, Not Just Scores

The company is not running purely on benchmarks. Spirit AI robots are operational on CATL battery production lines and deployed in JD Mall stores. It has secured a global strategic partnership with Bosch. According to the company, it has built a distributed real-world data collection network across more than 100 cities in China, with seven generations of proprietary wearable data collection hardware.

The plan is to accumulate millions of hours of real-world interaction data by end of 2026 — a direct response to Jensen Huang’s own framing of the problem at Computex: “For robotic systems and physical AI, data is the hardest problem.”

China’s Sweep of Physical AI Benchmarks

Spirit AI is not the only Chinese firm at the top of physical AI rankings. On WorldArena, which evaluates embodied world models, Manifold AI’s WorldScape-0.2 leads the policy evaluator track, ahead of Nvidia’s Cosmos-Predict 2.5. AgiBot leads the perception track. DexForce leads the data engine track.

Chinese firms attracted $3.4 billion in robotics venture funding in 2025, 42% more than the United States. Spirit AI’s result is evidence that the capital and the model quality are now moving together.

Nvidia remains the infrastructure layer: its GPUs train the models, its Jetson chips run them at the edge, and its Cosmos Coalition is designed to pull independent labs — including Chinese ones — onto its platform. That bet can coexist with losing a leaderboard title. Whether it survives losing industrial deployments is a different question.

Key Numbers

Model	Organization	RoboArena Score
Spirit v1.6	Spirit AI (Qianxun Intelligence)	1,924
Cosmos3-Nano-Policy	Nvidia	1,881
DreamZero	Nvidia-linked	1,763

• Spirit AI total raised in 3 months: ~$739M
• Valuation post-Series A+: $2.9B+
• Robots currently deployed: CATL battery lines, JD Mall stores
• Strategic partners: Bosch
• Investors include: Shunwei Capital, Yunfeng Capital, Sequoia China, Lei Jun, Jack Ma, 360, JD.com